Need is All You Need: Homeostatic Neural Networks Adapt to Concept Shift

TL;DR

The paper tackles non-stationary data by studying concept shift in online supervised learning and formalizes it as $P_{train}(y|x) \neq P_{test}(y|x)$. It introduces a homeostatic neural network in which the learner's vulnerability to inputs modulates its learning rate via excitatory/inhibitory effects, effectively putting its own 'skin in the game'. The authors demonstrate, on MNIST and Fashion-MNIST, that the homeostatic design yields superior adaptation to varying rates of concept shift, especially under high shift, compared with constant or randomly regulated learning rates. The results also show resilience to second-order shifts (seasonality) and reveal how the regulator tunes the learning rate to environmental dynamics. Limitations include use of only two datasets and a lack of benchmarking against certain momentum-based optimizers, pointing to future work on broader benchmarks and more systematic comparisons.

Abstract

In living organisms, homeostasis is the natural regulation of internal states aimed at maintaining conditions compatible with life. Typical artificial systems are not equipped with comparable regulatory features. Here, we introduce an artificial neural network that incorporates homeostatic features. Its own computing substrate is placed in a needful and vulnerable relation to the very objects over which it computes. For example, artificial neurons performing classification of MNIST digits or Fashion-MNIST articles of clothing may receive excitatory or inhibitory effects, which alter their own learning rate as a direct result of perceiving and classifying the digits. In this scenario, accurate recognition is desirable to the agent itself because it guides decisions to regulate its vulnerable internal states and functionality. Counterintuitively, the addition of vulnerability to a learner does not necessarily impair its performance. On the contrary, self-regulation in response to vulnerability confers benefits under certain conditions. We show that homeostatic design confers increased adaptability under concept shift, in which the relationships between labels and data change over time, and that the greatest advantages are obtained under the highest rates of shift. This necessitates the rapid un-learning of past associations and the re-learning of new ones. We also demonstrate the superior abilities of homeostatic learners in environments with dynamically changing rates of concept shift. Our homeostatic design exposes the artificial neural network's thinking machinery to the consequences of its own "thoughts", illustrating the advantage of putting one's own "skin in the game" to improve fluid intelligence.

Figures (5)

• Figure 1: Homeostatic learners incur some performance penalty in environments with no or low concept shift, but are far superior under conditions of highest shift. Color-coded validation accuracies of learners with their learning rates homeostatically regulated (blue), randomly regulated (green), and held constant (red). Traces show mean +/- SEM over 20 replicates.
• Figure 2: Accuracy and learning rate under “seasonality” of concept shift. Top row: Schedule A cycles between extreme rates of concept shift, while schedule B is more gradual. Middle row: The accuracy of the constant-LR classifier, in red, severely declines during stormy periods but returns to normal during calm periods. The homeostat, in blue, maintains good average performance across shifts in the rate of concept shift. Bottom row: The sequence of learning rates reveals that the homeostatic learner ratchets up its learning rate during stormy periods but is less inclined to reduce it during calm periods.
• Figure S1: Algorithm: Pseudocode for homeostatic self-regulation of learning rate.
• Figure S2: Concept shift is implemented by swapping the mapping between label and image between two randomly selected classes. Illustrated here on the MNIST dataset, the mappings for “zero” and “nine” are swapped. This swap will also invert the homeostatic effects expected for each number – a potentially hazardous situation for a vulnerable classifier.
• Figure S3: Learning rate sequencies of the two LR-regulating classifiers. The homeostatic learner seeks an LR appropriate to each level of concept shift, while the mean LR of the random regulator drifted upwards due to enforcement of a minimum non-zero LR. At 500 swaps per epoch (left, blue), the homeostat arrests its own LR growth and asymptotes. Data shown from MNIST only.