Loss of plasticity in deep continual learning

Specifics of continual backpropagation

Continual backpropagation selectively reinitializes low-utility units in the network. Our utility measure, called the contribution utility, is defined for each connection or weight and each unit. The basic intuition behind the contribution utility is that the magnitude of the product of units’ activation and outgoing weight gives information about how valuable this connection is to its consumers. If the contribution of a hidden unit to its consumer is small, its contribution can be overwhelmed by contributions from other hidden units. In such a case, the hidden unit is not useful to its consumer. We define the contribution utility of a hidden unit as the sum of the utilities of all its outgoing connections. The contribution utility is measured as a running average of instantaneous contributions with a decay rate, η, which is set to 0.99 in all experiments. In a feed-forward neural network, the contribution utility, u_l[i], of the ith hidden unit in layer l at time t is updated as

in which h_l,i,t is the output of the ith hidden unit in layer l at time t, w_l,i,k,t is the weight connecting the ith unit in layer l to the kth unit in layer l + 1 at time t and n_l+1 is the number of units in layer l + 1.

When a hidden unit is reinitialized, its outgoing weights are initialized to zero. Initializing the outgoing weights as zero ensures that the newly added hidden units do not affect the already learned function. However, initializing the outgoing weight to zero makes the new unit vulnerable to immediate reinitialization, as it has zero utility. To protect new units from immediate reinitialization, they are protected from a reinitialization for maturity threshold m number of updates. We call a unit mature if its age is more than m. Every step, a fraction of mature units ρ, called the replacement rate, is reinitialized in every layer.

The replacement rate ρ is typically set to a very small value, meaning that only one unit is replaced after hundreds of updates. For example, in class-incremental CIFAR-100 (Fig. 2) we used continual backpropagation with a replacement rate of 10^−5. The last layer of the network in that problem had 512 units. At each step, roughly 512 × 10^−5 = 0.00512 units are replaced. This corresponds to roughly one replacement after every 1/0.00512 ≈ 200 updates or one replacement after every eight epochs on the first five classes.

The final algorithm combines conventional backpropagation with selective reinitialization to continually inject random units from the initial distribution. Continual backpropagation performs a gradient descent and selective reinitialization step at each update. Algorithm 1 specifies continual backpropagation for a feed-forward neural network. In cases in which the learning system uses mini-batches, the instantaneous contribution utility can be used by averaging the utility over the mini-batch instead of keeping a running average to save computation (see Extended Data Fig. 5d for an example). Continual backpropagation overcomes the limitation of previous work^34,35 on selective reinitialization and makes it compatible with modern deep learning.

Algorithm 1

Continual backpropagation for a feed-forward network with L layers

Set replacement rate ρ, decay rate η and maturity threshold m

Initialize the weights w₀,…, w_L−1, in which w_l is sampled from distribution d_l

Initialize utilities u₁,…, u_L−1, number of units to replace c₁,…, c_L−1, and ages a₁,…, a_L−1 to 0

For each input x_t do

Forward pass: pass x_t through the network to get the prediction \(\widehat{{{\bf{y}}}_{t}}\)

Evaluate: receive loss \(l({{\bf{x}}}_{t},\widehat{{{\bf{y}}}_{t}})\)

Backward pass: update the weights using SGD or one of its variants

For layer l in 1: L − 1 do

Update age: a_l = a_l + 1

Update unit utility: see equation (1)

Find eligible units: n_eligible = number of units with age greater than m

Update number of units to replace: c_l = c_l + n_eligible × ρ

If c_l > 1

Find the unit with smallest utility and record its index as r

Reinitialize input weights: resample w_l−1[:,r] from distribution d_l

Reinitialize output weights: set w_l[r,:] to 0

Reinitialize utility and age: set u_l[r] = 0 and a_l[r] = 0

Update number of units to replace: c_l = c_l − 1

End For

End For

Details of Continual ImageNet

The ImageNet database we used consists of 1,000 classes, each of 700 images. The 700 images for each class were divided into 600 images for a training set and 100 images for a test set. On each binary classification task, the deep-learning network was first trained on the training set of 1,200 images and then its classification accuracy was measured on the test set of 200 images. The training consisted of several passes through the training set, called epochs. For each task, all learning algorithms performed 250 passes through the training set using mini-batches of size 100. All tasks used the downsampled 32 × 32 version of the ImageNet dataset, as is often done to save computation⁵¹.

All algorithms on Continual ImageNet used a convolutional network. The network had three convolutional-plus-max-pooling layers, followed by three fully connected layers, as detailed in Extended Data Table 3. The final layer consisted of just two units, the heads, corresponding to the two classes. At task changes, the input weights of the heads were reset to zero. Resetting the heads in this way can be viewed as introducing new heads for the new tasks. This resetting of the output weights is not ideal for studying plasticity, as the learning system gets access to privileged information on the timing of task changes (and we do not use it in other experiments in this paper). We use it here because it is the standard practice in deep continual learning for this type of problem in which the learning system has to learn a sequence of independent tasks⁵².

In this problem, we reset the head of the network at the beginning of each task. It means that, for a linear network, the whole network is reset. That is why the performance of a linear network will not degrade in Continual ImageNet. As the linear network is a baseline, having a low-variance estimate of its performance is desirable. The value of this baseline is obtained by averaging over thousands of tasks. This averaging gives us a much better estimate of its performance than other networks.

The network was trained using SGD with momentum on the cross-entropy loss and initialized once before the first task. The momentum hyperparameter was 0.9. We tested various step-size parameters for backpropagation but only presented the performance for step sizes 0.01, 0.001 and 0.0001 for clarity of Fig. 1b. We performed 30 runs for each hyperparameter value, varying the sequence of tasks and other randomness. Across different hyperparameters and algorithms, the same sequences of pairs of classes were used.

We now describe the hyperparameter selection for L2 regularization, Shrink and Perturb and continual backpropagation. The main text presents the results for these algorithms on Continual ImageNet in Fig. 1c. We performed a grid search for all algorithms to find the set of hyperparameters that had the highest average classification accuracy over 5,000 tasks. The values of hyperparameters used for the grid search are described in Extended Data Table 2. L2 regularization has two hyperparameters, step size and weight decay. Shrink and Perturb has three hyperparameters, step size, weight decay and noise variance. We swept over two hyperparameters of continual backpropagation: step size and replacement rate. The maturity threshold in continual backpropagation was set to 100. For both backpropagation and L2 regularization, the performance was poor for step sizes of 0.1 or 0.003. We chose to only use step sizes of 0.03 and 0.01 for continual backpropagation and Shrink and Perturb. We performed ten independent runs for all sets of hyperparameters. Then we performed another 20 runs to complete 30 runs for the best-performing set of hyperparameters to produce the results in Fig. 1c.

Class-incremental CIFAR-100

In the class-incremental CIFAR-100, the learning system gets access to more and more classes over time. Classes are provided to the learning system in increments of five. First, it has access to just five classes, then ten and so on, until it gets access to all 100 classes. The learning system is evaluated on the basis of how well it can discriminate between all the available classes at present. The dataset consists of 100 classes with 600 images each. The 600 images for each class were divided into 450 images to create a training set, 50 for a validation set and 100 for a test set. Note that the network is trained on all data from all classes available at present. First, it is trained on data from just five classes, then from all ten classes and so on, until finally, it is trained from data from all 100 classes simultaneously.

After each increment, the network was trained for 200 epochs, for a total of 4,000 epochs for all 20 increments. We used a learning-rate schedule that resets at the start of each increment. For the first 60 epochs of each increment, the learning rate was set to 0.1, then to 0.02 for the next 60 epochs, then 0.004 for the next 40 epochs and to 0.0008 for the last 40 epochs; we used the initial learning rate and learning-rate schedule reported in ref. ⁵³. During the 200 epochs of training for each increment, we kept track of the network with the best accuracy on the validation set. To prevent overfitting, at the start of each new increment, we reset the weights of the network to the weights of the best-performing (on the validation set) network found during the previous increment; this is equivalent to early stopping for each different increment.

We used an 18-layer deep residual network³⁸ for all experiments on class-incremental CIFAR-100. The network architecture is described in detail in Extended Data Table 1. The weights of convolutional and linear layers were initialized using Kaiming initialization⁵⁴, the weights for the batch-norm layers were initialized to one and all of the bias terms in the network were initialized to zero. Each time five new classes were made available to the network, five more output units were added to the final layer of the network. The weights and biases of these output units were initialized using the same initialization scheme. The weights of the network were optimized using SGD with a momentum of 0.9, a weight decay of 0.0005 and a mini-batch size of 90.

We used several steps of data preprocessing before the images were presented to the network. First, the value of all the pixels in each image was rescaled between 0 and 1 through division by 255. Then, each pixel in each channel was centred and rescaled by the average and standard deviation of the pixel values of each channel, respectively. Finally, we applied three random data transformations to each image before feeding it to the network: randomly horizontally flip the image with a probability of 0.5, randomly crop the image by padding the image with 4 pixels on each side and randomly cropping to the original size, and randomly rotate the image between 0 and 15°. The first two steps of preprocessing were applied to the training, validation and test sets, but the random transformations were only applied to the images in the training set.

We tested several hyperparameters to ensure the best performance for each different algorithm with our specific architecture. For the base system, we tested values for the weight decay parameter in {0.005, 0.0005, 0.00005}. A weight-decay value of 0.0005 resulted in the best performance in terms of area under the curve for accuracy on the test set over the 20 increments. For Shrink and Perturb, we used the weight-decay value of the base system and tested values for the standard deviation of the Gaussian noise in {10^−4, 10^−5, 10^−6}; 10^−5 resulted in the best performance. For continual backpropagation, we tested values for the maturity threshold in {1,000, 10,000} and for the replacement rate in {10^−4, 10^−5, 10^−6} using the contribution utility described in equation (1). A maturity threshold of 1,000 and a replacement rate of 10^−5 resulted in the best performance. Finally, for the head-resetting baseline, in Extended Data Fig. 1a, we used the same hyperparameters as for the base system, but the output layer was reinitialized at the start of each increment.

In Fig. 2d, we plot the stable rank of the representation in the penultimate layer of the network and the percentage of dead units in the full network. For a matrix \({\boldsymbol{\Phi }}\in {{\mathbb{R}}}^{n\times m}\) with singular values σ_k sorted in descending order for k = 1, 2,…, q and q = max(n, m), the stable rank⁵⁵ is \(\min \left\{k:\frac{{\Sigma }_{i}^{k}{\sigma }_{i}}{{\Sigma }_{j}^{q}{\sigma }_{j}} > 0.99\right\}\).

For reference, we also implemented a network with the same hyperparameters as the base system but that was reinitialized at the beginning of each increment. Figure 2b shows the performance of each algorithm relative to the performance of the reinitialized network. For completeness, Extended Data Fig. 1a shows the test accuracy of each algorithm in each different increment. The final accuracy of continual backpropagation on all 100 classes was 76.13%, whereas Extended Data Fig. 1b shows the performance of continual backpropagation for different replacement rates with a maturity threshold of 1,000. For all algorithms that we tested, there was no correlation between when a class was presented and the accuracy of that class, implying that the temporal order of classes did not affect performance.

Robust loss of plasticity in permuted MNIST

We now use a computationally cheap problem based on the MNIST dataset⁵⁶ to test the generality of loss of plasticity across various conditions. MNIST is one of the most common supervised-learning datasets used in deep learning. It consists of 60,000, 28 × 28, greyscale images of handwritten digits from 0 to 9, together with their correct labels. For example, the left image in Extended Data Fig. 3a shows an image that is labelled by the digit 7. The smaller number of classes and the simpler images enable much smaller networks to perform well on this dataset than are needed on ImageNet or CIFAR-100. The smaller networks in turn mean that much less computation is needed to perform the experiments and thus experiments can be performed in greater quantities and under a variety of different conditions, enabling us to perform deeper and more extensive studies of plasticity.

We created a continual supervised-learning problem using permuted MNIST datasets^57,58. An individual permuted MNIST dataset is created by permuting the pixels in the original MNIST dataset. The right image in Extended Data Fig. 3a is an example of such a permuted image. Given a way of permuting, all 60,000 images are permuted in the same way to produce the new permuted MNIST dataset. Furthermore, we normalized pixel values between 0 and 1 by dividing by 255.

By repeatedly randomly selecting from the approximately 10¹⁹³⁰ possible permutations, we created a sequence of 800 permuted MNIST datasets and supervised-learning tasks. For each task, we presented each of its 60,000 images one by one in random order to the learning network. Then we moved to the next permuted MNIST task and repeated the whole procedure, and so on for up to 800 tasks. No indication was given to the network at the time of task switching. With the pixels being permuted in a completely unrelated way, we might expect classification performance to fall substantially at the time of each task switch. Nevertheless, across tasks, there could be some savings, some improvement in speed of learning or, alternatively, there could be loss of plasticity—loss of the ability to learn across tasks. The network was trained on a single pass through the data and there were no mini-batches. We call this problem Online Permuted MNIST.

We applied feed-forward neural networks with three hidden layers to Online Permuted MNIST. We did not use convolutional layers, as they could not be helpful on the permuted problem because the spatial information is lost; in MNIST, convolutional layers are often not used even on the standard, non-permuted problem. For each example, the network estimated the probabilities of each of the tem classes, compared them to the correct label and performed SGD on the cross-entropy loss. As a measure of online performance, we recorded the percentage of times the network correctly classified each of the 60,000 images in the task. We plot this per-task performance measure versus task number in Extended Data Fig. 3b. The weights were initialized according to a Kaiming distribution.

The left panel of Extended Data Fig. 3b shows the progression of online performance across tasks for a network with 2,000 units per layer and various values of the step-size parameter. Note that that performance first increased across tasks, then began falling steadily across all subsequent tasks. This drop in performance means that the network is slowly losing the ability to learn from new tasks. This loss of plasticity is consistent with the loss of plasticity that we observed in ImageNet and CIFAR-100.

Next, we varied the network size. Instead of 2,000 units per layer, we tried 100, 1,000 and 10,000 units per layer. We ran this experiment for only 150 tasks, primarily because the largest network took much longer to run. The performances at good step sizes for each network size are shown in the middle panel of Extended Data Fig. 3b. Loss of plasticity with continued training is most pronounced at the smaller network sizes, but even the largest networks show some loss of plasticity.

Next, we studied the effect of the rate at which the task changed. Going back to the original network with 2,000-unit layers, instead of changing the permutation after each 60,000 examples, we now changed it after each 10,000, 100,000 or 1 million examples and ran for 48 million examples in total no matter how often the task changed. The examples in these cases were selected randomly with replacement for each task. As a performance measure of the network on a task, we used the percentage correct over all of the images in the task. The progression of performance is shown in the right panel in Extended Data Fig. 3b. Again, performance fell across tasks, even if the change was very infrequent. Altogether, these results show that the phenomenon of loss of plasticity robustly arises in this form of backpropagation. Loss of plasticity happens for a wide range of step sizes, rates of distribution change and for both underparameterized and overparameterized networks.

Loss of plasticity with different activations in the Slowly-Changing Regression problem

There remains the issue of the network’s activation function. In our experiments so far, we have used ReLU, the most popular choice at present, but there are several other possibilities. For these experiments, we switched to an even smaller, more idealized problem. Slowly-Changing Regression is a computationally inexpensive problem in which we can run a single experiment on a CPU core in 15 min, allowing us to perform extensive studies. As its name suggests, this problem is a regression problem—meaning that the labels are real numbers, with a squared loss, rather than nominal values with a cross-entropy loss—and the non-stationarity is slow and continual rather than abrupt, as in a switch from one task to another. In Slowly-Changing Regression, we study loss of plasticity for networks with six popular activation functions: sigmoid, tanh, ELU⁵⁹, leaky ReLU⁶⁰, ReLU⁶¹ and Swish⁶².

In Slowly-Changing Regression, the learner receives a sequence of examples. The input for each example is a binary vector of size m + 1. The input has f slow-changing bits, m − f random bits and then one constant bit. The first f bits in the input vector change slowly. After every T examples, one of the first f bits is chosen uniformly at random and its value is flipped. These first f bits remain fixed for the next T examples. The parameter T allows us to control the rate at which the input distribution changes. The next m − f bits are randomly sampled for each example. Last, the (m + 1)th bit is a bias term with a constant value of one.

The target output is generated by running the input vector through a neural network, which is set at the start of the experiment and kept fixed. As this network generates the target output and represents the desired solution, we call it the target network. The weights of the target networks are randomly chosen to be +1 or −1. The target network has one hidden layer with the linear threshold unit (LTU) activation. The value of the ith LTU is one if the input is above a threshold θ_i and 0 otherwise. The threshold θ_i is set to be equal to (m + 1) × β − S_i, in which β ∈ [0, 1] and S_i is the number of input weights with negative value⁶³. The details of the input and target function in the Slowly-Changing Regression problem are also described in Extended Data Fig. 2a.

The details of the specific instance of the Slowly-Changing Regression problem we use in this paper and the learning network used to predict its output are listed in Extended Data Table 4. Note that the target network is more complex than the learning network, as the target network is wider, with 100 hidden units, whereas the learner has just five hidden units. Thus, because the input distribution changes every T example and the target function is more complex than what the learner can represent, there is a need to track the best approximation.

We applied learning networks with different activation functions to the Slowly-Changing Regression. The learner used the backpropagation algorithm to learn the weights of the network. We used a uniform Kaiming distribution⁵⁴ to initialize the weights of the learning network. The distribution is U(−b, b) with bound, \(b={\rm{g}}{\rm{a}}{\rm{i}}{\rm{n}}\times \sqrt{\frac{3}{{\rm{n}}{\rm{u}}{\rm{m}}{\rm{\_}}{\rm{i}}{\rm{n}}{\rm{p}}{\rm{u}}{\rm{t}}{\rm{s}}}}\), in which gain is chosen such that the magnitude of inputs does not change across layers. For tanh, sigmoid, ReLU and leaky ReLU, the gain is 5/3, 1, \(\sqrt{2}\) and \(\sqrt{2/(1+{\alpha }^{2})}\), respectively. For ELU and Swish, we used \({\rm{gain}}=\sqrt{2}\), as was done in the original papers^59,62.

We ran the experiment on the Slowly-Changing Regression problem for 3 million examples. For each activation and value of step size, we performed 100 independent runs. First, we generated 100 sequences of examples (input–output pairs) for the 100 runs. Then these 100 sequences of examples were used for experiments with all activations and values of the step-size parameter. We used the same sequence of examples to control the randomness in the data stream across activations and step sizes.

The results of the experiments are shown in Extended Data Fig. 2b. We measured the squared error, that is, the square of the difference between the true target and the prediction made by the learning network. In Extended Data Fig. 2b, the squared error is presented in bins of 40,000 examples. This means that the first data point is the average squared error on the first 40,000 examples, the next is the average squared error on the next 40,000 examples and so on. The shaded region in the figure shows the standard error of the binned error.

Extended Data Fig. 2b shows that, in Slowly-Changing Regression, after performing well initially, the error increases for all step sizes and activations. For some activations such as ReLU and tanh, loss of plasticity is severe, and the error increases to the level of the linear baseline. Although for other activations such as ELU loss of plasticity is less severe, there is still a notable loss of plasticity. These results mean that loss of plasticity is not resolved by using commonly used activations. The results in this section complement the results in the rest of the article and add to the generality of loss of plasticity in deep learning.

Understanding loss of plasticity

We now turn our attention to understanding why backpropagation loses plasticity in continual-learning problems. The only difference in the learner over time is the network weights. In the beginning, the weights were small random numbers, as they were sampled from the initial distribution; however, after learning some tasks, the weights became optimized for the most recent task. Thus, the starting weights for the next task are qualitatively different from those for the first task. As this difference in the weights is the only difference in the learning algorithm over time, the initial weight distribution must have some unique properties that make backpropagation plastic in the beginning. The initial random distribution might have many properties that enable plasticity, such as the diversity of units, non-saturated units, small weight magnitude etc.

As we now demonstrate, many advantages of the initial distribution are lost concurrently with loss of plasticity. The loss of each of these advantages partially explains the degradation in performance that we have observed. We then provide arguments for how the loss of these advantages could contribute to loss of plasticity and measures that quantify the prevalence of each phenomenon. We provide an in-depth study of the Online Permuted MNIST problem that will serve as motivation for several solution methods that could mitigate loss of plasticity.

The first noticeable phenomenon that occurs concurrently with the loss of plasticity is the continual increase in the fraction of constant units. When a unit becomes constant, the gradients flowing back from the unit become zero or very close to zero. Zero gradients mean that the weights coming into the unit do not change, which means that this unit loses all of its plasticity. In the case of ReLU activations, this occurs when the output of the activations is zero for all examples of the task; such units are often said to be dead^64,65. In the case of the sigmoidal activation functions, this phenomenon occurs when the output of a unit is too close to either of the extreme values of the activation function; such units are often said to be saturated^66,67.

To measure the number of dead units in a network with ReLU activation, we count the number of units with a value of zero for all examples in a random sample of 2,000 images at the beginning of each new task. An analogous measure in the case of sigmoidal activations is the number of units that are ϵ away from either of the extreme values of the function for some small positive ϵ (ref. ⁶⁸). We only focus on ReLU networks in this section.

In our experiments on the Online Permuted MNIST problem, the deterioration of online performance is accompanied by a large increase in the number of dead units (left panel of Extended Data Fig. 3c). For the step size of 0.01, up to 25% of units die after 800 tasks. In the permuted MNIST problem, in which all inputs are positive because they are normalized between 0 and 1, once a unit in the first layer dies, it stays dead forever. Thus, an increase in dead units directly decreases the total capacity of the network. In the next section, we will see that methods that stop the units from dying can substantially reduce loss of plasticity. This further supports the idea that the increase in dead units is one of the causes of loss of plasticity in backpropagation.

Another phenomenon that occurs with loss of plasticity is the steady growth of the network’s average weight magnitude. We measure the average magnitude of the weights by adding up their absolute values and dividing by the total number of weights in the network. In the permuted MNIST experiment, the degradation of online classification accuracy of backpropagation observed in Extended Data Fig. 3b is associated with an increase in the average magnitude of the weights (centre panel of Extended Data Fig. 3c). The growth of the magnitude of the weights of the network can represent a problem because large weight magnitudes are often associated with slower learning. The weights of a neural network are directly linked to the condition number of the Hessian matrix in the second-order Taylor approximation of the loss function. The condition number of the Hessian is known to affect the speed of convergence of SGD algorithms (see ref. ⁶⁹ for an illustration of this phenomenon in convex optimization). Consequently, the growth in the magnitude of the weights could lead to an ill-conditioned Hessian matrix, resulting in a slower convergence.

The last phenomenon that occurs with the loss of plasticity is the drop in the effective rank of the representation. Similar to the rank of a matrix, which represents the number of linearly independent dimensions, the effective rank takes into consideration how each dimension influences the transformation induced by a matrix⁷⁰. A high effective rank indicates that most of the dimensions of the matrix contribute similarly to the transformation induced by the matrix. On the other hand, a low effective rank corresponds to most dimensions having no notable effect on the transformation, implying that the information in most of the dimensions is close to being redundant.

Formally, consider a matrix \(\Phi \in {{\mathbb{R}}}^{n\times m}\) with singular values σ_k for k = 1, 2,…, q, and q = max(n, m). Let p_k = σ_k/∥σ∥₁, in which σ is the vector containing all the singular values and ∥⋅∥₁ is the ℓ¹ norm. The effective rank of matrix Φ, or erank(Φ), is defined as

Note that the effective rank is a continuous measure that ranges between one and the rank of matrix Φ.

In the case of neural networks, the effective rank of a hidden layer measures the number of units that can produce the output of the layer. If a hidden layer has a low effective rank, then a small number of units can produce the output of the layer, meaning that many of the units in the hidden layer are not providing any useful information. We approximate the effective rank on a random sample of 2,000 examples before training on each task.

In our experiments, loss of plasticity is accompanied by a decrease in the average effective rank of the network (right panel of Extended Data Fig. 3c). This phenomenon in itself is not necessarily a problem. After all, it has been shown that gradient-based optimization seems to favour low-rank solutions through implicit regularization of the loss function or implicit minimization of the rank itself^71,72. However, a low-rank solution might be a bad starting point for learning from new observations because most of the hidden units provide little to no information.

The decrease in effective rank could explain the loss of plasticity in our experiments in the following way. After each task, the learning algorithm finds a low-rank solution for the current task, which then serves as the initialization for the next task. As the process continues, the effective rank of the representation layer keeps decreasing after each task, limiting the number of solutions that the network can represent immediately at the start of each new task.

In this section, we looked deeper at the networks that lost plasticity in the Online Permuted MNIST problem. We noted that the only difference in the learning algorithm over time is the weights of the network, which means that the initial weight distribution has some properties that allowed the learning algorithm to be plastic in the beginning. And as learning progressed, the weights of the network moved away from the initial distribution and the algorithm started to lose plasticity. We found that loss of plasticity is correlated with an increase in weight magnitude, a decrease in the effective rank of the representation and an increase in the fraction of dead units. Each of these correlates partially explains loss of plasticity faced by backpropagation.

Existing deep-learning methods for mitigating loss of plasticity

We now investigate several existing methods and test how they affect loss of plasticity. We study five existing methods: L2 regularization⁷³, Dropout⁷⁴, online normalization⁷⁵, Shrink and Perturb¹¹ and Adam⁴³. We chose L2 regularization, Dropout, normalization and Adam because these methods are commonly used in deep-learning practice. Although Shrink and Perturb is not a commonly used method, we chose it because it reduces the failure of pretraining, a problem that is an instance of loss of plasticity. To assess if these methods can mitigate loss of plasticity, we tested them on the Online Permuted MNIST problem using the same network architecture we used in the previous section, ‘Understanding loss of plasticity’. Similar to the previous section, we measure the online classification accuracy on all 60,000 examples of the task. All the algorithms used a step size of 0.003, which was the best-performing step size for backpropagation in the left panel of Extended Data Fig. 3b. We also use the three correlates of loss of plasticity found in the previous section to get a deeper understanding of the performance of these methods.

An intuitive way to address loss of plasticity is to use weight regularization, as loss of plasticity is associated with a growth of weight magnitudes, shown in the previous section. We used L2 regularization, which adds a penalty to the loss function proportional to the ℓ² norm of the weights of the network. The L2 regularization penalty incentivizes SGD to find solutions that have a low weight magnitude. This introduces a hyperparameter λ that modulates the contribution of the penalty term.

The purple line in the left panel of Extended Data Fig. 4a shows the performance of L2 regularization on the Online Permuted MNIST problem. The purple lines in the other panels of Extended Data Fig. 4a show the evolution of the three correlates of loss of plasticity with L2 regularization. For L2 regularization, the weight magnitude does not continually increase. Moreover, as expected, the non-increasing weight magnitude is associated with lower loss of plasticity. However, L2 regularization does not fully mitigate loss of plasticity. The other two correlates for loss of plasticity explain this, as the percentage of dead units kept increasing and the effective rank kept decreasing. Finally, Extended Data Fig. 4b shows the performance of L2 regularization for different values of λ. The regularization parameter λ controlled the peak of the performance and how quickly it decreased.

A method related to weight regularization is Shrink and Perturb¹¹. As the name suggests, Shrink and Perturb performs two operations; it shrinks all the weights and then adds random Gaussian noise to these weights. The introduction of noise introduces another hyperparameter, the standard deviation of the noise. Owing to the shrinking part of Shrink and Perturb, the algorithm favours solutions with smaller average weight magnitude than backpropagation. Moreover, the added noise prevents units from dying because it adds a non-zero probability that a dead unit will become active again. If Shrink and Perturb mitigates these correlates to loss of plasticity, it could reduce loss of plasticity.

The performance of Shrink and Perturb is shown in orange in Extended Data Fig. 4. Similar to L2 regularization, Shrink and Perturb stops the weight magnitude from continually increasing. Moreover, it also reduces the percentage of dead units. However, it has a lower effective rank than backpropagation, but still higher than that of L2 regularization. Not only does Shrink and Perturb have a lower loss of plasticity than backpropagation but it almost completely mitigates loss of plasticity in Online Permuted MNIST. However, Shrink and Perturb was sensitive to the standard deviation of the noise. If the noise was too high, loss of plasticity was much more severe, and if it was too low, it did not have any effect.

An important technique in modern deep learning is called Dropout⁷⁴. Dropout randomly sets each hidden unit to zero with a small probability, which is a hyperparameter of the algorithm. The performance of Dropout is shown in pink in Extended Data Fig. 4.

Dropout showed similar measures of percentage of dead units, weight magnitude and effective rank as backpropagation, but, surprisingly, showed higher loss of plasticity. The poor performance of Dropout is not explained by our three correlates of loss of plasticity, which means that there are other possible causes of loss of plasticity. A thorough investigation of Dropout is beyond the scope of this paper, though it would be an interesting direction for future work. We found that a higher Dropout probability corresponded to a faster and sharper drop in performance. Dropout with probability of 0.03 performed the best and its performance was almost identical to that of backpropagation. However, Extended Data Fig. 4a shows the performance for a Dropout probability of 0.1 because it is more representative of the values used in practice.

Another commonly used technique in deep learning is batch normalization⁷⁶. In batch normalization, the output of each hidden layer is normalized and rescaled using statistics computed from each mini-batch of data. We decided to include batch normalization in this investigation because it is a popular technique often used in practice. Because batch normalization is not amenable to the online setting used in the Online Permuted MNIST problem, we used online normalization⁷⁷ instead, an online variant of batch normalization. Online normalization introduces two hyperparameters used for the incremental estimation of the statistics in the normalization steps.

The performance of online normalization is shown in green in Extended Data Fig. 4. Online normalization had fewer dead units and a higher effective rank than backpropagation in the earlier tasks, but both measures deteriorated over time. In the later tasks, the network trained using online normalization has a higher percentage of dead units and a lower effective rank than the network trained using backpropagation. The online classification accuracy is consistent with these results. Initially, it has better classification accuracy, but later, its classification accuracy becomes lower than that of backpropagation. For online normalization, the hyperparameters changed when the performance of the method peaked, and it also slightly changed how fast it got to its peak performance.

No assessment of alternative methods can be complete without Adam⁴³, as it is considered one of the most useful tools in modern deep learning. The Adam optimizer is a variant of SGD that uses an estimate of the first moment of the gradient scaled inversely by an estimate of the second moment of the gradient to update the weights instead of directly using the gradient. Because of its widespread use and success in both supervised and reinforcement learning, we decided to include Adam in this investigation to see how it would affect the plasticity of deep neural networks. Adam has two hyperparameters that are used for computing the moving averages of the first and second moments of the gradient. We used the default values of these hyperparameters proposed in the original paper and tuned the step-size parameter.

The performance of Adam is shown in cyan in Extended Data Fig. 4. Adam’s loss of plasticity can be categorized as catastrophic, as it reduces substantially. Consistent with our previous results, Adam scores poorly in the three measures corresponding to the correlates of loss of plasticity. Adam had an early increase in the percentage of dead units that plateaus at around 60%, similar weight magnitude as backpropagation and a large drop in the effective rank early during training. We also tested Adam with different activation functions on the Slowly-Changing Regression and found that loss of plasticity with Adam is usually worse than with SGD.

Many of the standard methods substantially worsened loss of plasticity. The effect of Adam on the plasticity of the networks was particularly notable. Networks trained with Adam quickly lost almost all of their diversity, as measured by the effective rank, and gained a large percentage of dead units. This marked loss of plasticity of Adam is an important result for deep reinforcement learning, for which Adam is the default optimizer⁷⁸, and reinforcement learning is inherently continual owing to the ever-changing policy. Similar to Adam, other commonly used methods such as Dropout and normalization worsened loss of plasticity. Normalization had better performance in the beginning, but later it had a sharper drop in performance than backpropagation. In the experiment, Dropout simply made the performance worse. We saw that the higher the Dropout probability, the larger the loss of plasticity. These results mean that some of the most successful tools in deep learning do not work well in continual learning, and we need to focus on directly developing tools for continual learning.

We did find some success in maintaining plasticity in deep neural networks. L2 regularization and Shrink and Perturb reduce loss of plasticity. Shrink and Perturb is particularly effective, as it almost entirely mitigates loss of plasticity. However, both Shrink and Perturb and L2 regularization are slightly sensitive to hyperparameter values. Both methods only reduce loss of plasticity for a small range of hyperparameters, whereas for other hyperparameter values, they make loss of plasticity worse. This sensitivity to hyperparameters can limit the application of these methods to continual learning. Furthermore, Shrink and Perturb does not fully resolve the three correlates of loss of plasticity, it has a lower effective rank than backpropagation and it still has a high fraction of dead units.

We also applied continual backpropagation on Online Permuted MNIST. The replacement rate is the main hyperparameter in continual backpropagation, as it controls how rapidly units are reinitialized in the network. For example, a replacement rate of 10^−6 for our network with 2,000 hidden units in each layer would mean replacing one unit in each layer after every 500 examples.

Blue lines in Extended Data Fig. 4 show the performance of continual backpropagation. It has a non-degrading performance and is stable for a wide range of replacement rates. Continual backpropagation also mitigates all three correlates of loss of plasticity. It has almost no dead units, stops the network weights from growing and maintains a high effective rank across tasks. All algorithms that maintain a low weight magnitude also reduced loss of plasticity. This supports our claim that low weight magnitudes are important for maintaining plasticity. The algorithms that maintain low weight magnitudes were continual backpropagation, L2 regularization and Shrink and Perturb. Shrink and Perturb and continual backpropagation have an extra advantage over L2 regularization: they inject randomness into the network. This injection of randomness leads to a higher effective rank and lower number of dead units, which leads to both of these algorithms performing better than L2 regularization. However, continual backpropagation injects randomness selectively, effectively removing all dead units from the network and leading to a higher effective rank. This smaller number of dead units and a higher effective rank explains the better performance of continual backpropagation.

Details and further analysis in reinforcement learning

The experiments presented in the main text were conducted using the Ant-v3 environment from OpenAI Gym⁷⁹. We changed the coefficient of friction by sampling it log-uniformly from the range [0.02, 2.00], using a logarithm with base 10. The coefficient of friction changed at the first episode boundary after 2 million time steps had passed since the last change. We also tested Shrink and Perturb on this problem and found that it did not provide a marked performance improvement over L2 regularization. Two separate networks were used for the policy and the value function, and both had two hidden layers with 256 units. These networks were trained using Adam alongside PPO to update the weights in the network. See Extended Data Table 5 for the values of the other hyperparameters. In all of the plots showing results of reinforcement-learning experiments, the shaded region represents the 95% bootstrapped confidence⁸⁰.

The reward signal in the ant problem consists of four components. The main component rewards the agent for forward movement. It is proportional to the distance moved by the ant in the positive x direction since the last time step. The second component has a value of 1 at each time step. The third component penalizes the ant for taking large actions. This component is proportional to the square of the magnitude of the action. Finally, the last component penalizes the agent for large external contact forces. It is proportional to the sum of external forces (clipped in a range). The reward signal at each time step is the sum of these four components.

We also evaluated PPO and its variants in two more environments: Hopper-v3 and Walker-v3. The results for these experiments are presented in Extended Data Fig. 5a. The results mirrored those from Ant-v3; standard PPO suffered from a notable degradation in performance, in which its performance decreased substantially. However, this time, L2 regularization did not fix the issue in all cases; there was some performance degradation with L2 in Walker-v3. PPO, with continual backpropagation and L2 regularization, completely fixed the issue in all environments. Note that the only difference between our experiments and what is typically done in the literature is that we run the experiments for longer. Typically, these experiments are only done for 3 million steps, but we ran these experiments for up to 100 million steps.

PPO with L2 regularization only avoided degradation for a relatively large value of weight decay, 10^−3. This extreme regularization stops the agent from finding better policies and stays stuck at a suboptimal policy. There was large performance degradation for smaller values of weight decay, and for larger values, the performance was always low. When we used continual backpropagation and L2 regularization together, we could use smaller values of weight decay. All the results for PPO with continual backpropagation and L2 regularization have a weight decay of 10^−4, a replacement rate of 10^−4 and a maturity threshold of 10⁴. We found that the performance of PPO with continual backpropagation and L2 regularization was sensitive to the replacement rate but not to the maturity threshold and weight decay.

PPO uses the Adam optimizer, which keeps running estimates of the gradient and the squared of the gradient. These estimates require two further parameters, called β₁ and β₂. The standard values of β₁ and β₂ are 0.9 and 0.999, respectively, which we refer to as standard Adam. Lyle et al.²⁴ showed that the standard values of β₁ and β₂ cause a large loss of plasticity. This happens because of the mismatch in β₁ and β₂. A sudden large gradient can cause a very large update, as a large value of β₂ means that the running estimate for the square of the gradient, which is used in the denominator, is updated much more slowly than the running estimate for the gradient, which is the numerator. This loss of plasticity in Adam can be reduced by setting β₁ equal to β₂. In our experiments, we set β₁ and β₂ to 0.99 and refer to it as tuned Adam/PPO. In Extended Data Fig. 5c, we measure the largest total weight change in the network during a single update cycle for bins of 1 million steps. The first point in the plots shows the largest weight change in the first 1 million steps. The second point shows the largest weight change in the second 1 second steps and so on. The figure shows that standard Adam consistently causes very large updates to the weights, which can destabilize learning, whereas tuned Adam with β₁ = β₂ = 0.99 has substantially smaller updates, which leads to more stable learning. In all of our experiments, all algorithms other than the standard PPO used the tuned parameters for Adam (β₁ = β₂ = 0.99). The failure of standard Adam with PPO is similar to the failure of standard Adam in permuted MNIST.

In our next experiment, we perform a preliminary comparison with ReDo²⁵. ReDo is another selective reinitialization method that builds on continual backpropagation but uses a different measure of utility and strategy for reinitializing. We tested ReDo on Ant-v3, the hardest of the three environments. ReDo requires two parameters: a threshold and a reinitialization period. We tested ReDo for all combinations of thresholds in {0.01, 0.03, 0.1} and reinitialization periods in {10, 10², 10³, 10⁴, 10⁵}; a threshold of 0.1 with a reinitialization period of 10² performed the best. The performance of PPO with ReDo is plotted in Extended Data Fig. 5b. ReDo and continual backpropagation were used with weight decay of 10^−4 and β₁ and β₂ of 0.99. The figure shows that PPO with ReDo and L2 regularization performs much better than standard PPO. However, it still suffers from performance degradation and its performance is worse than PPO with L2 regularization. Note that this is only a preliminary comparison; we leave a full comparison and analysis of both methods for future work.

The performance drop of PPO in stationary environments is a nuanced phenomenon. Loss of plasticity and forgetting are both responsible for the observed degradation in performance. The degradation in performance implies that the agent forgot the good policy it had once learned, whereas the inability of the agent to relearn a good policy means it lost plasticity.

Loss of plasticity expresses itself in various forms in deep reinforcement learning. Some work found that deep reinforcement learning systems can lose their generalization abilities in the presence of non-stationarities⁸¹. A reduction in the effective rank, similar to the rank reduction in CIFAR-100, has been observed in some deep reinforcement-learning algorithms⁸². Nikishin et al.¹⁸ showed that many reinforcement-learning systems perform better if their network is occasionally reset to its naive initial state, retaining only the replay buffer. This is because the learning networks became worse than a reinitialized network at learning from new data. Recent work has improved performance in many reinforcement-learning problems by applying plasticity-preserving methods^{25,83,84,85,86,87}. These works focused on deep reinforcement learning systems that use large replay buffers. Our work complements this line of research as we studied systems based on PPO, which has much smaller replay buffers. Loss of plasticity is most relevant for systems that use small or no replay buffers, as large buffers can hide the effect of new data. Overcoming loss of plasticity is an important step towards deep reinforcement-learning systems that can learn from an online data stream.

Extended discussion

There are two main goals in continual learning: maintaining stability and maintaining plasticity^88,89,90,91. Maintaining stability is concerned with memorizing useful information and maintaining plasticity is about finding new useful information when the data distribution changes. Current deep-learning methods struggle to maintain stability as they tend to forget previously learned information^28,29. Many papers have been dedicated to maintaining stability in deep continual learning^{30,92,93,94,95,96,97}. We focused on continually finding useful information, not on remembering useful information. Our work on loss of plasticity is different but complementary to the work on maintaining stability. Continual backpropagation in its current form does not tackle the forgetting problem. Its current utility measure only considers the importance of units for current data. One idea to tackle forgetting is to use a long-term measure of utility that remembers which units were useful in the past. Developing methods that maintain both stability and plasticity is an important direction for future work.

There are many desirable properties for an efficient continual-learning system^98,99. It should be able to keep learning new things, control what it remembers and forgets, have good computational and memory efficiency and use previous knowledge to speed up learning on new data. The choice of the benchmark affects which property is being focused on. Most benchmarks and evaluations in our paper only focused on plasticity but not on other aspects, such as forgetting and speed of learning. For example, in Continual ImageNet, previous tasks are rarely repeated, which makes it effective for studying plasticity but not forgetting. In permuted MNIST, consecutive tasks are largely independent, which makes it suitable for studying plasticity in isolation. However, this independence means that previous knowledge cannot substantially speed up learning on new tasks. On the other hand, in class-incremental CIFAR-100, previous knowledge can substantially speed up learning of new classes. Overcoming loss of plasticity is an important, but still the first, step towards the goal of fast learning on future data^100,101,102. Once we have networks that maintain plasticity, we can develop methods that use previous knowledge to speed up learning on future data.

Loss of plasticity is a critical factor when learning continues for many tasks, but it might be less important if learning happens for a small number of tasks. Usually, the learning system can take advantage of previous learning in the first few tasks. For example, in class-incremental CIFAR-100 (Fig. 2), the base deep-learning systems performed better than the network trained from scratch for up to 40 classes. This result is consistent with deep-learning applications in which the learning system is first trained on a large dataset and then fine-tuned on a smaller, more relevant dataset. Plasticity-preserving methods such as continual backpropagation may still improve performance in such applications based on fine-turning, but we do not expect that improvement to be large, as learning happens only for a small number of tasks. We have observed that deep-learning systems gradually lose plasticity, and this effect accumulates over tasks. Loss of plasticity becomes an important factor when learning continues for a large number of tasks; in class-incremental CIFAR-100, the performance of the base deep-learning system was much worse after 100 classes.

We have made notable progress in understanding loss of plasticity. However, it remains unclear which specific properties of initialization with small random numbers are important for maintaining plasticity. Recent work^103,104 has made exciting progress in this direction and it remains an important avenue for future work. The type of loss of plasticity studied in this article is largely because of the loss of the ability to optimize new objectives. This is different from the type of loss of plasticity in which the system can keep optimizing new objectives but lose the ability to generalize^11,12. However, it is unclear if the two types of plasticity loss are fundamentally different or if the same mechanism can explain both phenomena. Future work that improves our understanding of plasticity and finds the underlying causes of both types of plasticity loss will be valuable to the community.

Continual backpropagation uses a utility measure to find and replace low-utility units. One limitation of continual backpropagation is that the utility measure is based on heuristics. Although it performs well, future work on more principled utility measures will improve the foundations of continual backpropagation. Our current utility measure is not a global measure of utility as it does not consider how a given unit affects the overall represented function. One possibility is to develop utility measures in which utility is propagated backwards from the loss function. The idea of utility in continual backpropagation is closely related to connection utility in the neural-network-pruning literature. Various papers^{105,106,107,108} have proposed different measures of connection utility for the network-pruning problem. Adapting these utility measures to mitigate loss of plasticity is a promising direction for new algorithms and some recent work is already making progress in this direction¹⁰⁹.

The idea of selective reinitialization is similar to the emerging idea of dynamic sparse training^110,111,112. In dynamic sparse training, a sparse network is trained from scratch and connections between different units are generated and removed during training. Removing connections requires a measure of utility, and the initialization of new connections requires a generator similar to selective reinitialization. The main difference between dynamic sparse training and continual backpropagation is that dynamic sparse training operates on connections between units, whereas continual backpropagation operates on units. Consequently, the generator in dynamic sparse training must also decide which new connections to grow. Dynamic sparse training has achieved promising results in supervised and reinforcement-learning problems^113,114,115, in which dynamic sparse networks achieve performance close to dense networks even at high sparsity levels. Dynamic sparse training is a promising idea that can be useful to maintain plasticity.

The idea of adding new units to neural networks is present in the continual-learning literature^92,116,117. This idea is usually manifested in algorithms that dynamically increase the size of the network. For example, one method¹¹⁷ expands the network by allocating a new subnetwork whenever there is a new task. These methods do not have an upper limit on memory requirements. Although these methods are related to the ideas in continual backpropagation, none are suitable for comparison, as continual backpropagation is designed for learning systems with finite memory, which are well suited for lifelong learning. And these methods would therefore require non-trivial modification to apply to our setting of finite memory.

Previous works on the importance of initialization have focused on finding the correct weight magnitude to initialize the weights. It has been shown that it is essential to initialize the weights so that the gradients do not become exponentially small in the initial layers of a network and the gradient is preserved across layers^54,66. Furthermore, initialization with small weights is critical for sigmoid activations as they may saturate if the weights are too large¹¹⁸. Despite all this work on the importance of initialization, the fact that its benefits are only present initially but not continually has been overlooked, as these papers focused on cases in which learning has to be done just once, not continually.

Continual backpropagation selectively reinitializes low-utility units. One common strategy to deal with non-stationary data streams is reinitializing the network entirely. In the Online Permuted MNIST experiment, full reinitialization corresponds to a performance that stays at the level of the first point (Extended Data Fig. 4a). In this case, continual backpropagation outperforms full reinitialization as it takes advantage of what it has previously learned to speed up learning on new data. In ImageNet experiments, the final performance of continual backpropagation is only slightly better than a fully reinitialized network (the first point for backpropagation in left panel of Fig. 1b). However, Fig. 1 does not show how fast an algorithm reaches the final performance in each task. We observed that continual backpropagation achieves the best accuracy ten times faster than a fully reinitialized network on the 5,000th task of Continual ImageNet, ten epochs versus about 125 epochs. Furthermore, continual backpropagation could be combined with other methods that mitigate forgetting, which can further speed up learning on new data. In reinforcement learning, full reinitialization is only practical for systems with a large buffer. For systems that keep a small or no buffer, such as those we studied, full reinitialization will lead the agent to forget everything it has learned, and its performance will be down to the starting point.

Loss of plasticity might also be connected to the lottery ticket hypothesis¹¹⁹. The hypothesis states that randomly initialized networks contain subnetworks that can achieve performance close to that of the original network with a similar number of updates. These subnetworks are called winning tickets. We found that, in continual-learning problems, the effective rank of the representation at the beginning of tasks reduces over time. In a sense, the network obtained after training on several tasks has less randomness and diversity than the original random network. The reduced randomness might mean that the network has fewer winning tickets. And this reduced number of winning tickets might explain loss of plasticity. Our understanding of loss of plasticity could be deepened by fully exploring its connection with the lottery ticket hypothesis.

Some recent works have focused on quickly adapting to the changes in the data stream^120,121,122. However, the problem settings in these papers were offline as they had two separate phases, one for learning and the other for evaluation. To use these methods online, they have to be pretrained on tasks that represent tasks that the learner will encounter during the online evaluation phase. This requirement of having access to representative tasks in the pretraining phase is not realistic for lifelong learning systems as the real world is non-stationary, and even the distribution of tasks can change over time. These methods are not comparable with those we studied in our work, as we studied fully online methods that do not require pretraining.

In this work, we found that methods that continually injected randomness while maintaining small weight magnitudes greatly reduced loss of plasticity. Many works have found that adding noise while training neural networks can improve training and testing performance. The main benefits of adding noise have been reported to be avoiding overfitting and improving training performance^123,124,125. However, it can be tricky to inject noise without degrading performance in some cases¹²⁶. In our case, when the data distribution is non-stationary, we found that continually injecting noise along with L2 regularization helps with maintaining plasticity in neural networks.