Deep learning provides new data-driven tools to relate neural activity to perception and cognition, aiding scientists in developing theories of neural computation that increasingly resemble biological systems both at the level of behavior and of neural activity. But what in a deep neural network should correspond to what in a biological system? This question is addressed implicitly in the use of comparison measures that relate specific neural or behavioral dimensions via a particular functional form. However, distinct comparison methodologies can give conflicting results in recovering even a known ground-truth model in an idealized setting, leaving open the question of what to conclude from the outcome of a systems comparison using any given methodology. Here, we develop a framework to make explicit and quantitative the effect of both hypothesis-driven aspects—such as details of the architecture of a deep neural network—as well as methodological choices in a systems comparison setting. We demonstrate via the learning dynamics of deep neural networks that, while the role of the comparison methodology is often de-emphasized relative to hypothesis-driven aspects, this choice can impact and even invert the conclusions to be drawn from a comparison between neural systems. We provide evidence that the right way to adjudicate a comparison depends on the use case—the scientific hypothesis under investigation—which could range from identifying single-neuron or circuit-level correspondences to capturing generalizability to new stimulus properties

Cognitive maps confer animals with flexible intelligence by representing spatial, temporal, and abstract relationships that can be used to shape thought, planning, and behavior. Cognitive maps have been observed in the hippocampus, but their algorithmic form and the processes by which they are learned remain obscure. Here, we employed large-scale, longitudinal two-photon calcium imaging to record activity from thousands of neurons in the CA1 region of the hippocampus while mice learned to efficiently collect rewards from two subtly different versions of linear tracks in virtual reality. The results provide a detailed view of the formation of a cognitive map in the hippocampus. Throughout learning, both the animal behavior and hippocampal neural activity progressed through multiple intermediate stages, gradually revealing improved task representation that mirrored improved behavioral efficiency. The learning process led to progressive decorrelations in initially similar hippocampal neural activity within and across tracks, ultimately resulting in orthogonalized representations resembling a state machine capturing the inherent struture of the task. We show that a Hidden Markov Model (HMM) and a biologically plausible recurrent neural network trained using Hebbian learning can both capture core aspects of the learning dynamics and the orthogonalized representational structure in neural activity. In contrast, we show that gradient-based learning of sequence models such as Long Short-Term Memory networks (LSTMs) and Transformers do not naturally produce such orthogonalized representations. We further demonstrate that mice exhibited adaptive behavior in novel task settings, with neural activity reflecting flexible deployment of the state machine. These findings shed light on the mathematical form of cognitive maps, the learning rules that sculpt them, and the algorithms that promote adaptive behavior in animals. The work thus charts a course toward a deeper understanding of biological intelligence and offers insights toward developing more robust learning algorithms in artificial intelligence.

What is the relationship between task, network architecture, and population activity in nonlinear deep networks? I will describe the Gated Deep Linear Network framework, which schematizes how pathways of information flow impact learning dynamics within an architecture. Because of the gating, these networks can compute nonlinear functions of their input. We derive an exact reduction and, for certain cases, exact solutions to the dynamics of learning. The reduction takes the form of a neural race with an implicit bias towards shared representations, which then govern the model’s ability to systematically generalize, multi-task, and transfer. We show how appropriate network architectures can help factorize and abstract knowledge. Together, these results begin to shed light on the links between architecture, learning dynamics and network performance.