Timothy Lillicrap
<p>April 1, 2019, 3:49 PM</p>
<p>talk at Cosyne in 2018.</p>
<p>The credit assignment problem. Must adjust earlier (upstream) weights to get right answer... and we use backprop. So, do we use a scalar or vector for weight adjustment.</p>
<p> </p>
<p>Could be simply a scalar that measures error (goal -result); </p>
<p> </p>
<p>Backprop underlies everything (unserpvised; supervised and RL);</p>
<p> </p>
<p>Why isn't backprop bio plausible. To do it we need... an error term; we need transpose of downstream weights; we need derivative of activation function; and separate forward and back passes.</p>
<p> </p>
<p>l </p>
<p> </p>
<p> </p>
<p> </p>
<p> </p>
<p> </p>
<p> </p>
<p> </p>
<p> </p>
<p> </p>
<p> </p>