Jan 8 2026

Things I Tried Today

Worked

• Shared initial layers & and shared loss in an actor critic model
• Additional simulation head in the AC net (big improvement in training speed in ablation study)
• Resetting weights after catastrophic forgetting
• Parquet for storing traces (20x more efficient than pickle)
• Tracking weight magnitude CDF (useful for debugging)
• Tracking overall + component losses over time
• Normalizing value net loss
• Using weight decay to reduce maximum weight magnitudes

Did Not Work

• Sine Layer: replace a fully connected linear layer with a fully connected sine layer, where each output is defined by sum[wi_a * sin(xi * wi_b)]
• Absolute Value Layer: replace a fully connected linear layer with a fully connected abs layer, where each output is defined by sum[wi * abs(xi)]

Meh

• Tracking individual parameter salience via integrated gradients. It wasn't very actionable. Salience may be more useful when tracked at a higher level, e.g. tracking the % of neurons in a layer that are salient. That might give some insight into the need for larger or smaller layers.