Autoshaped head poking in the mouse: a quantitative analysis of the learning curve

Abstract

In autoshaping experiments, we quantified the acquisition of anticipatory head poking in individual mice, using an algorithm that finds changes in the slope of a cumulative record. In most mice, upward changes in the amount of anticipatory poking per trial were abrupt, and tended to occur at session boundaries, suggesting that the session is as significant a unit of experience as the trial. There were large individual differences in the latency to the onset of vigorous responding. "Asymptotic" performance was unstable; large, bidirectional, and relatively enduring changes were common. Given the characteristics of the individual learning curves, it is unlikely that physiologically meaningful estimates of rate of learning can be extracted from group-average learning curves.

Fig 1. Raster plot of pokes during six sessions from a single mouse.

The thin horizontal lines represent single pokes and their length shows their duration. The heavy vertical line at 10 s indicates delivery of food.

Fig 2. Cumulative records of trial (solid line) and ITI (dashed line) poking proportions (left panel) and their corresponding selectivity scores (right panel) for six mice.

The vertical lines within the plots indicate session boundaries. Note the different scales for each mouse.

Fig 3. An illustrative cumulative record of 27 trials (taken with permission from Gallistel, Fairhurst, & Balsam, 2004).

No pecking occurred up to Trial 20. The change-point detection algorithm was applied iteratively to each successive point in the cumulative record, drawing in effect a straight line (not shown in the picture) from the point of origin up to that point. When it reached Trial 27, it drew the straight line indicated in the graph by the slanted dashed line. The point where the cumulative record deviates maximally from this straight line is between Trial 19 and 20. This point is identified as a putative change point. It divides the record into two portions: the cumulative pecks of the trials up to and including Trial 19, and the cumulative pecks of Trials 20 to 27. The distribution of pecks per trials for the trials before and after the putative change point are compared, using an appropriate statistical test. If they differ significantly, the putative change point is accepted as valid. Then the algorithm begins over again, taking Trial 19 as the origin (zero point of the cumulative record) and the pecks on Trial 20 as the first datum (first measurement in the postchange cumulative record).

Fig 4. Upper panels: Cumulative records of the selectivity scores for the same mouse (M339) shown in Figure 1, with the change points indicated by the superposed circles.

Lower panels: Step plots of the slopes of the cumulative records between successive change points. On the left, a liberal criterion was used in the finding of change points (logit  =  2); on the right, a conservative criterion (logit  =  6) was used.

Fig 5. Acquisition profile histograms.

The bar graphs show the number of mice that exhibited the particular type of acquisition profile that is displayed below each bar.

Fig 6. Slope plots generated by the algorithm when applied to the cumulative record of selectivity scores averaged across all mice.

Only the first 464 trials for each subject were included in the averaging process.

Fig 7. Histograms of the distribution of selectivity-score change points within a session for the mice of all groups (change points from all sessions have been pooled in the histograms).

The upper panel shows all change points. The middle and lower panels show upward and downward change points respectively. The left panels show Criterion-2 analyses, and the right panels show Criterion-6 analyses.

Fig 8. Selectivity score slope plots from selected mice, generated by the algorithm with Criterion 6.

Notice the instability in postacquisition performance.

Fig 9. Upper graph: mean number of trials per session for each group (1/4d > 1/2d > 1/d & 2/d, p < .001).

Bottom graph: mean selectivity score during asymptote for each group (1/4d > 1/d, p  =  .04). The error bars represent standard error of the means. Significant differences are denoted with an asterisk.

Fig 10. Sessions to onset of above-baseline conditioned responding (upper graph) and to the trial that led to asymptotic performance (lower graph) for each group.

Fig 11. Slope plots of the cumulative logs of trial-initiation latencies for two mice, as reported by the algorithm with Criterion 6.

The arithmetic means of the log trial-start latencies of each segment of the cumulative records (that is, the slopes of these plots) have been converted into seconds and plotted on the y-axis as geometric means of the raw trial-start latencies.

Fig 12. Similar behavioral effects of changing growth function and performance function.

On the right (Underlying Function) are two models (M1 & M2) of associative strength growth as a function of trials. The plot on the left (Performance Function, rotated 90° counter clockwise, because its abscissa is the ordinate of the Underlying Function plot) shows two instances (P1 & P2) of behavioral response growth as a function of associative strength. The dashed lines show the limits of a performance window, formed by a threshold for observable behavior and a ceiling at which behavior saturates. The latency measures along the trial axis on the right (L_1,1, etc) are the number of trials to the abrupt appearance of conditioned behavior for different combinations of model and performance functions.