Existing machines are functionally specific tools that were made for easy
prediction and control. Tomorrow’s machines may be closer to biological systems
in their mutability, resilience, and autonomy. But first they must be capable
of learning and retaining new information without being exposed to it
arbitrarily often. Past efforts to engineer such systems have sought to build
or regulate artificial neural networks using disjoint sets of weights that are
uniquely sensitive to specific tasks or inputs. This has not yet enabled
continual learning over long sequences of previously unseen data without
corrupting existing knowledge: a problem known as catastrophic forgetting. In
this paper, we introduce a system that can learn sequentially over previously
unseen datasets (ImageNet, CIFAR-100) with little forgetting over time. This is
done by controlling the activity of weights in a convolutional neural network
on the basis of inputs using top-down regulation generated by a second
feed-forward neural network. We find that our method learns continually under
domain transfer with sparse bursts of activity in weights that are recycled
across tasks, rather than by maintaining task-specific modules. Sparse synaptic
bursting is found to balance activity and suppression such that new functions
can be learned without corrupting extant knowledge, thus mirroring the balance
of order and disorder in systems at the edge of chaos. This behavior emerges
during a prior pre-training (or ‘meta-learning’) phase in which regulated
synapses are selectively disinhibited, or grown, from an initial state of
uniform suppression through prediction error minimization.
Today’s machines are designed for easy prediction and control, but the machines of the future may resemble biological systems in their mutability, resilience, and autonomy. However, in order to achieve this, these machines need to be capable of learning and retaining new information without being exposed to it repeatedly. The problem of catastrophic forgetting, which occurs when existing knowledge is corrupted by new information, has hindered previous efforts to engineer such systems.
In this paper, the authors introduce a system that can learn sequentially over previously unseen datasets with little forgetting over time. This is achieved by controlling the activity of weights in a convolutional neural network using top-down regulation generated by a second feed-forward neural network. By utilizing sparse bursts of activity in weights that are recycled across tasks, rather than maintaining task-specific modules, the method enables continual learning under domain transfer.
The key to this approach is the concept of sparse synaptic bursting, which balances activity and suppression in the neural network. This allows new functions to be learned without corrupting existing knowledge. This phenomenon mirrors the balance of order and disorder in natural systems at the edge of chaos.
This behavior emerges during a prior pre-training or “meta-learning” phase, in which regulated synapses are selectively disinhibited or grown from an initial state of uniform suppression through prediction error minimization. It is through this process that the system can learn sequentially over previously unseen datasets without catastrophic forgetting.
The concepts discussed in this paper highlight the multi-disciplinary nature of building machine learning systems that can achieve continual learning. By incorporating elements from computer science (neural networks), biology (biological systems), and chaos theory (balance of order and disorder), the authors have developed a promising approach for future machines.