Poster #70 - A machine-learning approach to the development of problem solving in infant locomotion

• In Event: 2-002 - Special Poster Session 05 with Continental Breakfast Reception
  In Poster Session: PS 05 - Interdisciplinary Section

Integrative Statement

Goal-directed locomotion involves problem solving (Adolph & Robinson, 2015)—how to descend a steep slope, navigate a drop-off, or cross a narrow bridge. Some locomotor solutions require a particular coordination among arms and legs. For example, “cruising” infants must coordinate arm and leg movements to move sideways in a upright posture while holding onto a supporting surface to augment balance. Previous research showed that infants successfully adapt their movements to cruise over gaps in a handrail (Adolph et al 2010).
In the current study, we focused on how infants successfully solve the problem of cruising over gaps. We observed 24 11-month-old cruising infants (13 boys) on an adjustable “gaps” apparatus (Adolph et al 2010). The apparatus had an adjustable gap in the handrail that infants held for support. Cruising over the gap required infants to position their hands and legs on the near side of the gap so they could grasp the handrail on the far side of the gap without losing balance. Each trial began with infants standing in a sideways position, holding the handrail on the starting platform. To objectively identify infants’ problem-solving strategies, we developed a unique machine-learning process (Figure 1) that includes: (1) a computer vision algorithm to identify distances between limbs; (2) a dynamic time warping algorithm to identify similarities in real-time changes of limb movements across trials and infants, and (3) density-peaks clustering to detect different crossing strategies.
Machine-learning analyses revealed eight different crossing strategies (Figure 1 step 6, Figure 2A). We validated the strategies with visual inspection of the video for each trial. We found that more experienced cruisers converged on a 1:1 ratio between the number of strategies they used and the number of different gap sizes they crossed successfully (Figure 2B-C). That is, experienced cruisers switched strategies only when it was required—they used the same strategy for the same gap size, and different strategies for different gap sizes. Novice cruisers, in contrast, either used multiple strategies to cross the same gap, or consistently used the same strategy regardless of gap size. These results indicate that locomotor experience improves the outcome (the proportion of successful trials), and does so by changing the process by which infants accomplish the goal.
The current study suggests that a process-oriented approach can illuminate our understanding about the development of problem solving. The use of an innovative combination of methods demonstrates how machine learning can inform us about the “how” of problem solving and offers new tools for studying real-time infant behavior.

Authors

• Ori Ossmy, New York University
  Presenting Author

• Karen E Adolph, New York University
  Non-Presenting Author