Churchland lab

A recent paper in Neuron from Kanaka Rajan, Chris Harvey and David Tank sets out to demonstrate how relatively unstructured networks can give rise to highly structured outputs that persist on slow timescales relevant to behaviors like decision-making and working memory. Such unstructured networks seem at first like exactly the wrong thing to support stimulus-driven persistent activity. Indeed, classic work in the prefrontal cortex revealed individual neurons that respond persistently during delays, presumably support the ability of the animal to hold information in mind over that delay. In mouse posterior parietal cortex, however, it’s a different story. On a memory guided decision task published previously many individual neurons respond only very transiently, for much less time than the animal holds those memories in mind. Both that paper and the current one argue that many such neurons could fire in sequence, supporting slow-timescale memory-guided decisions even in the absence of single neurons with persistent activity.

The big steps forward in the current paper are:

1. The authors demonstrated that a randomly connected network could give rise to this activity. This was an advance for a number of reasons, including the development of a new model framework called PINning. This method builds on a now classic technique, FORCE learning which generates coherent activity patterns from chaotic networks. PINing is different because only a small percentage (~12%) of synaptic weights are allowed to change. The ability of the network to capture the complex firing rates of 437 neurons when only a few synaptic weights were allowed to change is a big deal.
   

   Network that learns by PINning; red lines are the only synapses that are allowed to change during learning to match the data.

2. The paper pointed out features of the data that are incompatible with a traditional model for persistent activity, like bump attractors. This is evidence against an appealing idea (that may be present in other systems) in which a hill of activity moves around the network, driving a persistent response.
3. Finally, the authors found that the network’s success relied not only on the strongly choice-selective neurons you might expect, but also on neurons that weren’t selective for the animal’s choice at all. In fact, they observed that these seemingly unimportant neurons might play a critical “conveyer belt” role that was essential in supporting more difficult decisions, especially those among many alternatives. The previous paper (and indeed many other studies) mainly excluded these neurons from analysis; an understandable choice at the time, but one that now warrants reconsideration.

There is still a challenge ahead for putative mechanisms that support slow timescale behaviors like working memory and decision-making. At the moment, there are few causal manipulations that can disrupt proposed mechanisms and demonstrate an effect on behavior. In the framework here, it would be compelling to demonstrate that changing the order of the sequence changed the behavior (admittedly no small feat!). More traditional mechanisms aren’t off the hook either: demonstrating that persistent activity at the single neuron level supports working memory likewise would be aided by precise disruption experiments. Indeed, the single neuron persistence could be epiphenomenological; the persistent working memories could be supported by some other aspect of the network. Many such  manipulation experiments will be feasible in the near future.

Until then, I am excited to see a new mechanism to support slow-timescale behavior. It is counterintuitive that such network complexity can be captured by a randomly connected network, especially one in which such a small number of synapses are allowed to change.