- Can fish learn to discriminate between complex auditory stimuli that humans can also learn to discriminate?
- If so, does this behavior generalize to novel auditory stimuli satisfying the same discriminability?
As the complex auditory stimuli, Chase uses musical audio. The stimuli used in three experiments are differentiable by humans along ``genre'' or ``style'', i.e., ``blues'' music (recordings of music by John Lee Hooker, Muddy Waters, and Blues compilations) and ``classical'' music (recordings of music by Bach, Handel, Vivaldi, etc.). The stimuli used in the fourth experiment are differentiable by humans along melody. The results of the first three experiments show three fish capable in learning to discriminate between "blues" and "classical" music. Furthermore, probes with novel stimuli, and an iterative reversal test, argue that the fish were "exhibiting open-ended categorization" of the two kinds of stimuli.
The first two experiments do not show what features of the stimulus are being used by the fish, or whether they are performing the task "on the basis of deeper generic attributes", rather than an unknown discriminant. The third experiment thus attempts to answer whether the fish can learn to discriminate between musical sounds that are labeled either "blues" or "classical" without the cue of timbre. Music by John Lee Hooker, Bach, and Vivaldi were transcribed to MIDI, and rendered with the same and/or different instruments. The results show the fish learn to discriminate between the "styles", even when the signals are produced with the same MIDI instrument; but that performance decreases when the same instruments are used then when they are different. The final experiment attempts to answer whether the fish can learn to discriminate between two different melodies though they have the same rhythm and timbre. The stimuli are a melody of Paganini, and one where the notes are reversed, but the timing is preserved. The one fish completing the experiment shows a capacity to perform this task. Several trials controlled for local features, such as starting and ending notes. Further work is necessary to determine what features the fish are learning to use to perform this task; but the experiments definitely prove fish are able to work with complex auditory stimuli.
I really like how this paper shows the time and effort necessary to scientifically and efficiently answer real questions --- even though it treats musical genre in an artificial way (i.e., Aristotelian categorization). (To criticize this point any further will miss the message of the work.) The experiments it discusses take place over what appears to be at least two years, and which can't really take place over any shorter time-span because living subjects can only be rewarded with so much food at a time. Such a time commitment absolutely requires a solid experimental plan to minimize waste and maximize results. Chase performs four experiments (and a preliminary one), but answers several questions with each experiment, and also uses elements of one experiment to prepare the subjects for the next experiment. The work in this paper exemplifies the kinds of considerations taken for granted in much experimental work in my own disciplines (signal processing and machine learning applied to audio and music signals). Machines need no reward or reinforcement, but why should they be evaluated any differently than Chase evaluates Beauty, Oro and Pepi (the three fish)?