We　present　an　integrated　model　for　Bayesian　learning　of　sparse　representation　and　classifier　training,　and　apply　it　for　the　task　of　visual　recognition.　Most　previous　work　learns　the　sparse　representation　and　trains　the　classifier　on　top　of　it　in　two　separate　steps.　We　cast　these　two　into　a　unified　probabilistic　model.　This　way,　the　supervised　labels　can　effectively　affect　the　learning　of　the　sparse　representation.　In　the　training　phase,　the　inference　of　the　joint　expectation　for　dictionary,　code,　classifier　and　other　variables　under　the　observation　of　descriptors　and　labels　is　carried　out　by　Gibbs　Sampling.　In　the　testing　phase,　based　on　the　learned　parameters,　the　sparse　code　and　the　class　label　of　the　image　are　obtained　by　Bayesian　inference.　The　proposed　model　is　evaluated　on　Caltech　101　dataset　and　its　efficacy　is　demonstrated　by　a　careful　analysis　of　the　experimental　results.　©　Springer　International　Publishing　Switzerland　2013.